Inkling 2 Small: Thinking Machines Unveils Powerhouse AI with 594B Parameters
Mira Murati's Thinking Machines unveils Inkling 2 Small, a 594B parameter model excelling in agentic coding and vision.
The artificial intelligence landscape has a new heavyweight contender. Thinking Machines Lab, the ambitious startup founded by former OpenAI executive Mira Murati, has officially announced the launch of Inkling 2 Small. This second-generation model represents a massive leap in capability, designed to push the boundaries of autonomous workflows and complex reasoning.
The Architecture: Massive Scale, Efficient Execution
At first glance, the "Small" in the name might be misleading. Inkling 2 Small is built on a Mixture-of-Experts (MoE) architecture boasting a staggering 594 billion total parameters. However, to maintain efficiency during actual computation, the model only activates 17 billion parameters at any given time.
This strategic design allows the model to handle immense amounts of data and complex tasks without requiring the computational overhead of a fully dense model of the same size.
Breaking Performance Records
Thinking Machines reports that Inkling 2 Small has significantly outperformed its predecessors. According to the company's blog, the model achieved a score of 31 on the Artificial Analysis Intelligence Index. These gains are most evident in four key areas:
- Agentic Coding: Superior ability to write and refine code autonomously.
- Vision: The capacity to process and understand image inputs.
- Knowledge Work: Enhanced handling of information-dense tasks.
- Reasoning: Improved logical deduction and problem-solving.
Notably, the model is designed to handle variable computational effort, though the company clarified that audio support is currently excluded from this release.
Mastering the 'Agentic' Workflow
One of the most impressive claims made by Thinking Machines is the model's ability to operate within "agent loops." Because it can incorporate visual feedback, Inkling 2 Small can iteratively build and refine 3D designs, games, and websites with interactive elements.
The company attributes this success to a rigorous training process that included:
- More efficient pre-training methods.
- A massive, unified Reinforcement Learning (RL) run involving hundreds of thousands of long-horizon coding tasks.
- The use of synthetic coding pipelines and agentic RL evaluators that analyzed not just the final code, but the method used to solve the problem.
Before its official launch, preliminary versions of the system were tested on the OpenCode benchmark under the name Fledge-Alpha to ensure it could handle real-world developer challenges.
Honest Admissions: The Limitations
Despite the breakthroughs, Thinking Machines has been transparent about the model's current shortcomings. Inkling 2 Small is not yet fully optimized for conversational interfaces. Users may encounter:
- Issues with textual formatting and an unusual tendency to generate excessive forward-slashes.
- Verbose phrasing in responses.
- Occasional errors in tool calls when working with software environments that have very large toolsets.
- Increased sensitivity regarding user safety and mental health requests.
Availability and the Road to Open Weights
Inkling 2 Small is initially launching with closed weights and will be accessible through Tinker. However, the company is committed to a "safe path to open weights." Thinking Machines intends to release the full weights under an Apache 2.0 license next week, following the completion of external red-teaming and safety testing.
A Billion-Dollar Foundation
The launch comes on the heels of a massive financial injection. In July 2025, Thinking Machines raised USD 2 billion in an early-stage funding round, catapulting the company's valuation to USD 12 billion. This funding was backed by some of the biggest names in the tech and investment worlds, including Nvidia, AMD, Cisco, and Jane Street.